[Serusers] assertion failed in t_retransmit_reply; race condition in t_newtran?

[Valentin Nechayev]

Hi,

we had got series of SER cores with failed assertion in
t_retransmit_reply():

        assert(t->uas.response.dst.send_sock);

Essential backtrace part:

#0  0x48167fc4 in kill () from /usr/lib/libc.so.4
#1  0x481a993e in abort () from /usr/lib/libc.so.4
#2  0x481858d3 in __assert () from /usr/lib/libc.so.4
#3  0x48281da3 in t_retransmit_reply () from /usr/local/lib/ser/modules/tm.so
#4  0x48277433 in t_newtran () from /usr/local/lib/ser/modules/tm.so

This place from ser.cfg:

        if (!t_newtran()) {
                sl_send_reply("500", "could not create transaction");
                break;
        };

response.dst.send_sock is filled in init_rb(), which is called
from t_newtran(). t_newtran() has the following:

=== cut ===
        UNLOCK_HASH(p_msg->hash_index);
        /* now, when the transaction state exists, check if
           there is a meaningful Via and calculate it; better
           do it now than later: state is established so that
           subsequent retransmissions will be absorbed and will
          not possibly block during Via DNS resolution; doing
           it later would only burn more CPU as if there is an
           error, we cannot relay later whatever comes out of the
           the transaction
        */
        if (!init_rb( &T->uas.response, p_msg)) {
                LOG(L_ERR, "ERROR: t_newtran: unresolvable via1\n");
                put_on_wait( T );
                t_unref(p_msg);
                return E_BAD_VIA;
        }

        return 1;
=== end cut ===

Does there race condition between UNLOCK_HASH(), init_rb() and
parallel t_newtran() from another SER process exist?

Details: FreeBSD 4.10, SER 0.9.3 (with PortaOne additions, but not
essential for the question).

-- 
Valentin Nechayev
PortaOne Inc., Software Engineer
mailto:netch at portaone.com